CN117237286A - Method for detecting internal defects of gas-insulated switchgear - Google Patents
Method for detecting internal defects of gas-insulated switchgear Download PDFInfo
- Publication number
- CN117237286A CN117237286A CN202311128356.9A CN202311128356A CN117237286A CN 117237286 A CN117237286 A CN 117237286A CN 202311128356 A CN202311128356 A CN 202311128356A CN 117237286 A CN117237286 A CN 117237286A
- Authority
- CN
- China
- Prior art keywords
- module
- convolution
- feature map
- network
- mask
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000007547 defect Effects 0.000 title claims abstract description 64
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000001514 detection method Methods 0.000 claims abstract description 66
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 42
- 238000012549 training Methods 0.000 claims abstract description 7
- 238000000605 extraction Methods 0.000 claims description 80
- 230000004927 fusion Effects 0.000 claims description 55
- 238000010586 diagram Methods 0.000 claims description 21
- 238000009826 distribution Methods 0.000 claims description 17
- 238000010606 normalization Methods 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 11
- 238000005070 sampling Methods 0.000 claims description 6
- 230000003628 erosive effect Effects 0.000 abstract 1
- 230000006870 function Effects 0.000 description 7
- 238000002679 ablation Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 238000004590 computer program Methods 0.000 description 5
- 238000004422 calculation algorithm Methods 0.000 description 4
- 230000004913 activation Effects 0.000 description 3
- 238000004458 analytical method Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 238000002474 experimental method Methods 0.000 description 3
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 102100031315 AP-2 complex subunit mu Human genes 0.000 description 1
- 101000796047 Homo sapiens AP-2 complex subunit mu Proteins 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000004140 cleaning Methods 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 238000000354 decomposition reaction Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 239000003292 glue Substances 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 239000012212 insulator Substances 0.000 description 1
- 230000009191 jumping Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 239000002184 metal Substances 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000012360 testing method Methods 0.000 description 1
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y04—INFORMATION OR COMMUNICATION TECHNOLOGIES HAVING AN IMPACT ON OTHER TECHNOLOGY AREAS
- Y04S—SYSTEMS INTEGRATING TECHNOLOGIES RELATED TO POWER NETWORK OPERATION, COMMUNICATION OR INFORMATION TECHNOLOGIES FOR IMPROVING THE ELECTRICAL POWER GENERATION, TRANSMISSION, DISTRIBUTION, MANAGEMENT OR USAGE, i.e. SMART GRIDS
- Y04S10/00—Systems supporting electrical power generation, transmission or distribution
- Y04S10/50—Systems or methods supporting the power network operation or management, involving a certain degree of interaction with the load-side end user applications
Abstract
The invention discloses a method for detecting internal defects of gas-insulated switchgear, which comprises the steps of acquiring an internal environment label image of the gas-insulated switchgear containing the defects and an internal environment non-label image of the gas-insulated switchgear; constructing REP-YOLOX target detection model based on convolutional neural network and transducer network; training a REP-YOLOX target detection model based on a convolutional neural network and a Transformer network by using an internal environment label image of the gas-insulated switchgear containing the defects; and performing defect detection on the non-tag image of the internal environment of the gas-insulated switchgear by using a trained REP-YOLOX target detection model based on a convolutional neural network and a Transformer network. The invention can obviously improve the detection rate of defects such as the erosion points, the foreign matters and the like in the gas-insulated switchgear.
Description
Technical Field
The invention belongs to the technical field of equipment defect detection, relates to internal defect detection of a gas-insulated switchgear, and particularly relates to the technical field of defect detection of gas-insulated switchgear (Gas insulated switchgear, GIS).
Background
The GIS equipment has the advantages of high safety, strong reliability, small occupied area, strong anti-interference performance and the like, is an indispensable part in a power transmission system, and is also an important guarantee for maintaining the normal operation of a power grid. The GIS equipment is externally wrapped by a metal cavity, and the inside of the GIS equipment is composed of a lightning arrester, a supporting insulator and a part of a grounding switch. However, in the process of installing and overhauling GIS equipment, foreign matters such as screws and nuts are easy to leave, and in addition, in the long-term operation of the GIS equipment, discharge is easy to occur, so that ablation defects are caused, and the normal operation of the GIS equipment is affected, and even some harm which is difficult to measure is caused.
The internal space of GIS equipment is complicated and narrow and small, and maintenance personnel can hardly enter the equipment for checking and cleaning. The traditional GIS equipment internal defect detection method mainly comprises a partial discharge detection method and an SF6 decomposition substance detection method, but the methods all need expensive detection equipment, require extremely high proficiency of operators and are easy to cause the problem of missed detection.
At present, the feature extraction and analysis of the GIS equipment internal defect detection algorithm based on the traditional image processing are very complex, and the generalization capability of the algorithms is not strong; the detection algorithm based on the traditional deep learning cannot efficiently analyze global information, so that the problems of insufficient accuracy, poor robustness, slower detection speed, easiness in interference of illumination information and the like of the internal defect detection of GIS equipment are caused.
Disclosure of Invention
Aiming at the problems of poor robustness, low detection speed and the like in the existing GIS defect detection technology at present, the invention provides a method for detecting the internal defects of gas-insulated switchgear, which effectively reduces the calculated amount by carrying out feature extraction through convolutional neural networks (Convolutional Neural Network, CNN) and can effectively realize the detection of the internal defects of the GIS based on the depth feature weighting of a transducer.
In order to achieve the above purpose, the present invention is realized by adopting the following technical scheme.
The invention provides a method for detecting internal defects of gas-insulated switchgear, which comprises the following steps:
s1, acquiring an internal environment label image of gas insulated switchgear containing defects and an internal environment non-label image of the gas insulated switchgear;
s2, constructing a REP-YOLOX target detection model based on a convolutional neural network and a transducer network; the REP-YOLOX target detection model based on the convolutional neural network and the transducer network specifically comprises the following steps: a backbone feature extraction network based on a convolutional neural network, a neck network based on a transducer network and a prediction network; the trunk feature extraction network based on the convolutional neural network comprises a plurality of feature extraction layers in series, wherein the feature extraction layers are combined with cavity convolution to sequentially extract features of an input image from shallow to deep, and an extracted feature map is input into a neck network; the neck network based on the Transformer network is combined with mask convolution to shade the input feature map according to a distribution rule, feature extraction under different shading conditions is completed, and the extracted feature map is input into a prediction network; obtaining a defect identification result through the prediction network;
s3, training a REP-YOLOX target detection model based on a convolutional neural network and a transducer network by using an internal environment label image of the gas-insulated switchgear containing the defects;
s4, performing defect detection on the non-tag image of the internal environment of the gas-insulated switchgear by using a trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network.
Further, the backbone feature extraction network based on the convolutional neural network specifically comprises: an upper feature extraction layer, a middle and lower feature extraction layer and a bottom feature extraction layer;
the upper layer feature extraction layer comprises a Focus module, a first convolution module and an enhanced CSPLlayer module which are connected in sequence; after inputting image data to the Focus module, outputting enhanced convolution characteristics to an intermediate characteristic extraction layer through a first convolution module and an enhanced CSPLlayer module;
the intermediate feature extraction layer comprises a second convolution module and a first CSPLlayer module which are sequentially connected; after the enhanced convolution feature is input to the second convolution module, a first feature map is output to a middle-lower feature extraction layer and a neck network based on a transform network through a first CSPLayer module;
the middle and lower feature extraction layer comprises a third convolution module and a second CSPLlayer module which are sequentially connected; inputting the first feature map to the third convolution module, and outputting a second feature map to a bottom feature extraction layer and a neck network based on a transform network through a second CSPLayer module;
the bottom layer feature extraction layer comprises a fourth convolution module, an SPPNet module and a third CSPLlayer module which are sequentially connected; and the fourth convolution module inputs the second feature map and outputs a third feature map to the neck network based on the transform network through the SPPNet module and the third CSPLlayer module.
Further, the enhanced CSPLlayer module comprises a first hole convolution module, a second hole convolution module, a residual error module and a third hole convolution module; and the feature map input to the second cavity convolution module is processed by a residual error module, is spliced with the feature map output by the first cavity convolution module, and is further subjected to feature extraction by a third cavity convolution module.
Further, the SPPNet module comprises an input convolution module, three convolution layers with different kernel sizes and an output convolution module, wherein the three convolution layers are arranged in parallel; the three convolution layers with different kernel sizes are arranged in parallel, and the feature images extracted by the convolution layers are spliced with the feature images output by the input convolution module and then output by the output convolution module.
Further, the neck network based on the Transformer network specifically includes:
the device comprises a convolution layer, a fourth CSPLAyer module, a first mask convolution module, a second mask convolution module, a fifth CSPLAyer module, a sixth CSPLAyer module, a seventh CSPLAyer module and a transducer module;
inputting a third characteristic diagram into the convolution layer, and obtaining a fourth characteristic diagram after convolution operation;
splicing the fourth feature map with a second feature map input into a fourth CSPLayer module after up-sampling operation, fusing feature map information and further extracting features to obtain a first fused feature map;
inputting a first fusion feature map to the first mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a first mask feature map;
inputting a first feature map to the second mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a second mask feature map;
the first mask feature map is spliced with the second mask feature map after up-sampling operation, feature map information is fused, a second fused feature map is further obtained through feature extraction of a fifth CSPLayer module, and the second fused feature map is output to a prediction network;
the second fusion feature map is spliced with the first mask feature map after downsampling operation, feature map information is fused, a third fusion feature map is further obtained through feature extraction of a sixth CSPLlayer module, and the third fusion feature map is output to a prediction network;
the third fusion feature map is spliced with the fourth feature map after downsampling operation, feature map information is fused, the fourth fusion feature map is further obtained through feature extraction of a seventh CSPLayer module, and the fourth fusion feature map is output to a transducer module;
and the converter module performs global feature extraction on the fourth fusion feature map to obtain a fifth fusion feature map, and outputs the fifth fusion feature map to the prediction network.
Further, the first mask convolution module and the second mask convolution module have the same structure, and specifically include:
a first mask convolution layer, a second mask convolution layer, a third mask convolution layer, a fourth mask convolution layer, a fifth mask convolution layer, and a sixth mask convolution layer in parallel;
the first mask convolution layer processes the input first fusion feature map by adopting a blank first mask, and then performs feature extraction through the convolution layer and the normalization layer to obtain a first mask convolution feature map;
the second to sixth mask convolution layers are used for processing the input first fusion feature images by adopting the second to sixth masks which are shielded according to a distribution rule, and feature extraction is performed through the convolution layers and the normalization layers to obtain second to sixth mask convolution feature images;
further, the transducer module specifically includes:
a convolutional layer a, transformer network, convolutional layer b, and convolutional layer c;
the convolution layer a performs position coding on the input fourth fusion feature map, performs global feature extraction through a Transformer network, and performs convolution operation through the convolution layer b to obtain a global feature map;
and splicing the global feature map with the fourth fusion feature map through jump connection, and then performing convolution operation through a convolution layer c to obtain a fifth fusion feature map.
Further, in the trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network, a first mask convolution module is subjected to model re-parameters, an input first fusion feature map is overlapped according to the channel dimension, and six convolutions of the first mask convolution module are overlapped according to the channel dimension.
Further, in the trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network, a second mask convolution module is subjected to model re-parameters, an input first feature map is overlapped according to the channel dimension, and six convolutions of the second mask convolution module are overlapped according to the channel dimension.
The invention has the following beneficial effects:
the invention improves the receptive field of the backbone network by effectively combining the hollow convolution and the conventional convolution, and can effectively reduce the calculated amount; then, the feature extraction capability under the condition of shielding different defects is improved by using a mask convolution module, and the detection speed is ensured by using a model heavy parameter technology; and finally, carrying out global feature analysis on a prediction part of the model by utilizing a transducer module, and effectively realizing detection of defects (especially, ablation points, screws and the like) in GIS equipment by means of depth feature weighting of the transducer, thereby obviously improving the detection rate of defects such as the ablation points, foreign matters and the like in the gas-insulated switchgear.
Drawings
Fig. 1 is a schematic flow chart of a method for detecting internal defects of a gas-insulated switchgear according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a REP-YOLOX target detection model based on a convolutional neural network and a transducer network according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a conventional convolution (a), a hole convolution (b), and an enhanced CSPLlayer module (c) according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a mask convolution module according to an embodiment of the present disclosure;
FIG. 5 is a schematic diagram of a transducer module according to an embodiment of the present invention;
FIG. 6 is a graph showing the AP50 values of different types of defects according to an embodiment of the present invention;
fig. 7 is a schematic diagram of a defect detection experiment result of a GIS device in an embodiment of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and all the inventions which make use of the inventive concept are protected by the spirit and scope of the present invention as defined and defined in the appended claims to those skilled in the art.
Examples
As shown in fig. 1, the method for detecting internal defects of a gas-insulated switchgear according to an embodiment of the present invention includes steps S1 to S4 as follows:
s1, acquiring an internal environment label image of gas insulated switchgear containing defects and an internal environment non-label image of the gas insulated switchgear;
in an optional embodiment of the present invention, the present embodiment first obtains an internal environment image of a gas-insulated switchgear containing a defect, then marks the defect in the internal environment image of the gas-insulated switchgear to obtain an internal defect image of the gas-insulated switchgear with a mark and a tag, and finally divides the internal defect image of the gas-insulated switchgear with the mark and the tag into a training set and a verification set, for performing model training on the REP-YOLOX target detection model based on the convolutional neural network and the transform network constructed in step S2.
In addition, the embodiment further performs defect detection by acquiring a non-tag image of the internal environment of the gas-insulated switchgear to be detected and using the REP-YOLOX target detection model based on the convolutional neural network and the transducer network trained in the step S3.
S2, constructing a REP-YOLOX target detection model based on a convolutional neural network and a transducer network;
in an optional embodiment of the present invention, the REP-YOLOX target detection model based on the convolutional neural network and the transducer network constructed in this embodiment specifically includes:
a convolutional neural network-based backbone feature extraction network, a transducer network-based neck network, and a prediction network.
The following describes in detail the respective partial networks of the REP-YOLOX target detection model based on convolutional neural network and transducer network constructed according to the present invention with reference to FIG. 2, respectively.
The trunk feature extraction network based on the convolutional neural network in this embodiment specifically includes:
an upper feature extraction layer, a middle and lower feature extraction layer and a bottom feature extraction layer;
the upper layer feature extraction layer comprises a Focus module, a first convolution module and an enhanced CSPLlayer module which are connected in sequence; after inputting image data to the Focus module, outputting enhanced convolution characteristics to an intermediate characteristic extraction layer through a first convolution module and an enhanced CSPLlayer module;
the intermediate feature extraction layer comprises a second convolution module and a first CSPLlayer module which are sequentially connected; after the enhanced convolution feature is input to the second convolution module, a first feature map is output to a middle-lower feature extraction layer and a neck network based on a transform network through a first CSPLayer module;
the middle and lower feature extraction layer comprises a third convolution module and a second CSPLlayer module which are sequentially connected; inputting the first feature map to the third convolution module, and outputting a second feature map to a bottom feature extraction layer and a neck network based on a transform network through a second CSPLayer module;
the bottom layer feature extraction layer comprises a fourth convolution module, an SPPNet module and a third CSPLlayer module which are sequentially connected; and outputting a third characteristic diagram to a neck network based on a transform network through an SPPNet module and a third CSPLlayer module after inputting the second characteristic diagram to the fourth convolution module.
The enhanced CSPLlayer module comprises a first cavity convolution module, a second cavity convolution module, a residual error module and a third cavity convolution module; and the feature map input to the second cavity convolution module is processed by a residual error module, is spliced with the feature map output by the first cavity convolution module, and is further subjected to feature extraction by a third cavity convolution module. The first hole convolution module, the second hole convolution module and the third hole convolution module have the same structure and comprise a hole convolution layer (convolution kernel size 3*3, hole rate is 2), a BN layer and an activation function (Silu) which are arranged in series. The residual error module comprises a fifth convolution module and a sixth convolution module which are arranged in series, and the input end of the residual error module is connected to the output end in a jumping mode.
The first CSPLAyer module, the second CSPLAyer module, the third CSPLAyer module, the fourth CSPLAyer module, the fifth CSPLAyer module, the sixth CSPLAyer module and the seventh CSPLAyer module have the same structures and comprise a seventh convolution module, an eighth convolution module, a residual error module and a ninth convolution module; and the feature map input to the seventh convolution module is processed by a residual error module, is spliced with the feature map output by the eighth convolution module, and is further subjected to feature extraction by the ninth convolution module.
The SPPNet module comprises an input convolution module, three convolution layers with different kernel sizes (the convolution kernel sizes are 5*5, 9*9 and 13 x 13 respectively) which are arranged in parallel, and an output convolution module; the three convolution layers with different kernel sizes are arranged in parallel, and the feature images are spliced with the input convolution module and then output through the output convolution module.
The first convolution module, the second convolution module, the third convolution module, the fourth convolution module, the sixth convolution module, the seventh convolution module, the eighth convolution module, the ninth convolution module, and the tenth, eleventh, twelfth, thirteenth, fourteenth, input and output convolution modules all have the same structure, and comprise serially arranged convolution layers (the convolution kernel size is 3*3), BN layers and an activation function (Silu). The fifth convolution module includes a convolution layer (convolution kernel size 1*1), a BN layer, and an activation function (Silu).
In the target detection task, the larger receptive field can be helpful for assisting global features, and has a larger influence on detection accuracy. In order to reduce parameters of the model, the existing target detection algorithm adopts pooling operation to reduce the size of the input image and increase the receptive field, so that certain information is lost, and the omission of defects is particularly easy to cause. In order to solve this problem, the embodiment promotes the receptive field of the backbone network by effectively combining the hole convolution and the conventional convolution, where the hole convolution is implemented by adding holes in the feature map based on the conventional convolution, as shown in fig. 3.
The neck network based on the transform network in this embodiment specifically includes:
the device comprises a convolution layer, a fourth CSPLAyer module, a first mask convolution module, a second mask convolution module, a fifth CSPLAyer module, a sixth CSPLAyer module, a seventh CSPLAyer module and a transducer module;
inputting a third characteristic diagram into the first convolution module, and obtaining a fourth characteristic diagram after convolution operation;
the fourth feature map is spliced with a second feature map input into a fourth CSPLayer module after up-sampling operation, feature map information is fused, and further features are extracted to obtain a first fused feature map;
inputting a first fusion feature map to the first mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a first mask feature map;
inputting a first feature map to the second mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a second mask feature map;
the first mask feature map is spliced with the second mask feature map after up-sampling operation, feature map information is fused, a second fused feature map is further obtained through feature extraction of a fifth CSPLayer module, and the second fused feature map is output to a prediction network;
the second fusion feature map is spliced with the first mask feature map after downsampling operation, feature map information is fused, a third fusion feature map is further obtained through feature extraction of a sixth CSPLlayer module, and the third fusion feature map is output to a prediction network;
the third fusion feature map is spliced with the fourth feature map after downsampling operation, feature map information is fused, the fourth fusion feature map is further obtained through feature extraction of a seventh CSPLayer module, and the fourth fusion feature map is output to a transducer module;
and the converter module performs global feature extraction on the fourth fusion feature map to obtain a fifth fusion feature map, and outputs the fifth fusion feature map to the prediction network.
The first mask convolution module and the second mask convolution module have the same structure, and specifically comprise:
the first, second, third, fourth, fifth and sixth mask convolution layers are in parallel. The second mask to the fifth mask are equally divided into four equal parts, each mask shields an area, the sixth mask shields a central area, and the shielding area is the same as the shielding area of other masks.
The first mask convolution layer processes the input first fusion feature map by adopting a blank first mask, and then performs feature extraction through the convolution layer and the normalization layer to obtain a first mask convolution feature map;
the second mask convolution layer processes the input first fusion feature map by adopting a second mask (a lower right area is blocked) which is blocked according to a distribution rule, and then the feature extraction is carried out through the convolution layer and the normalization layer to obtain a second mask convolution feature map;
the third mask convolution layer processes the input first fusion feature map by adopting a third mask (covering the upper left area) which is shielded according to a distribution rule, and then performs feature extraction through the convolution layer and the normalization layer to obtain a third mask convolution feature map;
the fourth mask convolution layer processes the input first fusion feature map by adopting a fourth mask (covering an upper right area) which is shielded according to a distribution rule, and then performs feature extraction through the convolution layer and the normalization layer to obtain a fourth mask convolution feature map;
the fifth mask convolution layer processes the input first fusion feature map by adopting a fifth mask (shielding lower left area) which is shielded according to a distribution rule, and then performs feature extraction through the convolution layer and the normalization layer to obtain a fifth mask convolution feature map;
the sixth mask convolution layer processes the input first fusion feature map by adopting a sixth mask (a shielding central area) which is shielded according to a distribution rule, and then performs feature extraction through the convolution layer and the normalization layer to obtain a sixth mask convolution feature map;
and adding the first mask convolution feature map, the second mask convolution feature map, the third mask convolution feature map, the fourth mask convolution feature map, the fifth mask convolution feature map and the sixth mask convolution feature map to obtain a first mask feature map.
The second mask convolution module operates the same as the first mask convolution film, and has the input of the first feature map and the output of the second mask feature map.
The GIS equipment has the defect dense distribution condition, and different degrees of shielding can also appear among different defects, so that a certain difficulty is brought to the feature extraction of the existing CNN model. In order to solve this problem, the Mask convolution (Mask Conv) module is used to improve the capability of extracting the defect features under different shielding conditions, and compared with the conventional CNN module, five parallel CNN modules are added to the Mask convolution to extract features of feature graphs which are shielded to a certain extent according to the distribution rule, and finally output features are added, as shown in fig. 4. The embodiment improves the feature extraction capability under the condition of shielding different defects by using the mask convolution module, and ensures the detection speed by using a model heavy parameter technology.
The transducer module in this embodiment specifically includes:
convolution layer a (convolution kernel size 3*3), a Transformer network, convolution layer b (convolution kernel size 1*1), and convolution layer c (convolution kernel size 1*1);
the convolution layer a performs position coding on the input fourth fusion feature map, performs global feature extraction through a Transformer network, and performs convolution operation through the convolution layer b to obtain a global feature map;
and splicing the global feature map with the fourth fusion feature map through jump connection, and then performing convolution operation through a convolution layer c to obtain a fifth fusion feature map.
In the defect detection task, the global feature is very important, and the identification rate of the defects can be improved through analysis of the defects and the surrounding environment. In the embodiment, the transducer module is adopted to extract global features, so that the detection accuracy of the model is further improved. The structure of the transducer module is shown in fig. 5, wherein the transducer network adopts a conventional arrangement, and mainly comprises two linear normalization layers, a multi-layer perceptron layer and a self-care layer, and the two linear normalization layers in the transducer network are mainly used for accelerating the calculation speed of the model and improving the stability of the model. The invention further increases the stability of the model by adding a 3 x 3 convolutional layer a as the position code of the transducer module and by a jump connection.
In the embodiment, three yolhead of the prediction network are decoupled and output according to the second fusion feature map, the third fusion feature map and the fifth fusion feature map respectively to obtain three category prediction results Cls, target frame prediction results Reg and target existence probabilities Obj under different sizes, and then three prediction parameters under the same size are overlapped to obtain corresponding prediction results; the final detection result is the union of three YoloHead prediction results.
The three YoloHead structures are the same, and as shown in FIG. 2, the YoloHead structures comprise a tenth convolution module, an eleventh convolution module, a twelfth convolution module, a thirteenth convolution module and a fourteenth convolution module; the eleventh convolution module and the thirteenth convolution module are respectively connected with the output end of the tenth convolution module; the eleventh convolution module and the twelfth convolution module are arranged in series and output a category prediction result Cls; the thirteenth convolution module and the fourteenth convolution module are arranged in series, and output a target frame prediction result Reg and a target existence probability Obj.
S3, training a REP-YOLOX target detection model based on a convolutional neural network and a transducer network by using an internal environment label image of the gas-insulated switchgear containing the defects.
In an alternative embodiment of the present invention, the present embodiment uses the training set and the verification set that divide the image of the internal environment label of the gas-insulated switchgear including the defect in step S1 to train the REP-YOLOX target detection model based on the convolutional neural network and the Transformer network constructed in step S2, and continuously performs iterative optimization on the predicted weight through forward propagation and backward propagation, so as to obtain the trained REP-YOLOX target detection model based on the convolutional neural network and the Transformer network.
S4, performing defect detection on the internal environment image of the gas-insulated switchgear by using a trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network.
In an optional embodiment of the present invention, in the trained REP-YOLOX target detection model based on the convolutional neural network and the transform network, the first mask convolution module performs model weight parameters, the input first fusion feature map is overlapped according to the channel dimension, and six convolutions of the first mask convolution module are overlapped according to the channel dimension.
And in the trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network, performing model re-parameters on the second mask convolution module, superposing the input first feature map according to the channel dimension, and superposing six convolutions of the second mask convolution module according to the channel dimension.
In order to improve the detection speed of the model, the method carries out model weight parameters on the trained mask convolution module, superimposes input features according to the channel dimension, and superimposes six convolutions of the mask convolution module according to the channel dimension, so that the mask convolution module can be simplified into a CNN module in an inference stage, and the detection speed of the model is improved on the premise of ensuring the feature extraction capability.
As shown in Table 1, the invention analyzes typical defect experimental results in four GIS devices, namely, a corrosion point, a nut, a rubber and a screw. Because the screw, the nut and the glue have obvious characteristics and are relatively unobvious to strong light reflection, the invention provides the AP of the model for the three defects 50 、F1 50 The accuracy and recall rate can all reach the highest level of the dataset. As the defect of the ablation point has larger reflection under strong light, the recall rate of the model is influenced, but the invention provides the model AP 50 96.58% can still be achieved, and the specific class AP graph is shown in fig. 6. Meanwhile, in order to more intuitively display the model detection result, the invention visualizes part of the test set, as shown in fig. 7.
Table 1: GIS equipment defect detection result table
To verify the effectiveness of the proposed method, four ablation experiments were designed as shown in table 2. The receptive field of the network can be improved after the cavity convolution is added into the backbone network, and the detection accuracy of the model is slightly improved. Then, after the Mask Conv module is used, the feature extraction capability of the model for shielding defects in different degrees is enhanced, and the detection precision of the model, in particular mAP, is greatly improved 75 The lifting rate is 12.91 percent. Finally, the method uses the transducer module to extract global characteristics, so that the detection accuracy of the model is further improved, and the mAP50 value reaches 99.14%.
Table 2: ablation experiment result table
The present invention is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.
These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.
The principles and embodiments of the present invention have been described in detail with reference to specific examples, which are provided to facilitate understanding of the method and core ideas of the present invention; meanwhile, as those skilled in the art will have variations in the specific embodiments and application scope in accordance with the ideas of the present invention, the present description should not be construed as limiting the present invention in view of the above.
Those of ordinary skill in the art will recognize that the embodiments described herein are for the purpose of aiding the reader in understanding the principles of the present invention and should be understood that the scope of the invention is not limited to such specific statements and embodiments. Those of ordinary skill in the art can make various other specific modifications and combinations from the teachings of the present disclosure without departing from the spirit thereof, and such modifications and combinations remain within the scope of the present disclosure.
Claims (9)
1. A method for detecting internal defects of a gas-insulated switchgear, comprising the steps of:
s1, acquiring an internal environment label image of gas insulated switchgear containing defects and an internal environment non-label image of the gas insulated switchgear;
s2, constructing a REP-YOLOX target detection model based on a convolutional neural network and a transducer network; the REP-YOLOX target detection model based on the convolutional neural network and the transducer network specifically comprises the following steps: a backbone feature extraction network based on a convolutional neural network, a neck network based on a transducer network and a prediction network; the trunk feature extraction network based on the convolutional neural network comprises a plurality of feature extraction layers in series, wherein the feature extraction layers are combined with cavity convolution to sequentially extract features of an input image from shallow to deep, and an extracted feature map is input into a neck network; the neck network based on the Transformer network is combined with mask convolution to shade the input feature map according to a distribution rule, feature extraction under different shading conditions is completed, and the extracted feature map is input into a prediction network; obtaining a defect identification result through the prediction network;
s3, training a REP-YOLOX target detection model based on a convolutional neural network and a transducer network by using an internal environment label image of the gas-insulated switchgear containing the defects;
s4, performing defect detection on the non-tag image of the internal environment of the gas-insulated switchgear by using a trained REP-YOLOX target detection model based on the convolutional neural network and the transducer network.
2. The method for detecting internal defects of a gas-insulated switchgear according to claim 1, wherein the convolutional neural network-based trunk feature extraction network specifically comprises: an upper feature extraction layer, a middle and lower feature extraction layer and a bottom feature extraction layer;
the upper layer feature extraction layer comprises a Focus module, a first convolution module and an enhanced CSPLlayer module which are connected in sequence; after inputting image data to the Focus module, outputting enhanced convolution characteristics to an intermediate characteristic extraction layer through a first convolution module and an enhanced CSPLayer module;
the intermediate feature extraction layer comprises a second convolution module and a first CSPLlayer module which are sequentially connected; after the enhanced convolution feature is input to the second convolution module, a first feature map is output to a middle-lower feature extraction layer and a neck network based on a transform network through a first CSPLayer module;
the middle and lower feature extraction layer comprises a third convolution module and a second CSPLlayer module which are sequentially connected; inputting the first feature map to the third convolution module, and outputting a second feature map to a bottom feature extraction layer and a neck network based on a transform network through a second CSPLayer module;
the bottom layer feature extraction layer comprises a fourth convolution module, an SPPNet module and a third CSPLlayer module which are sequentially connected; and outputting a third characteristic diagram to a neck network based on a transform network through an SPPNet module and a third CSPLlayer module after inputting the second characteristic diagram to the fourth convolution module.
3. The method for detecting internal defects of a gas-insulated switchgear according to claim 2, wherein the enhanced CSPLayer module comprises a first hole convolution module, a second hole convolution module, a residual module, and a third hole convolution module; and the feature map input to the second cavity convolution module is processed by a residual error module, is spliced with the feature map output by the first cavity convolution module, and is further subjected to feature extraction by a third cavity convolution module.
4. The method for detecting internal defects of a gas-insulated switchgear according to claim 2, wherein the SPPNet module comprises an input convolution module, three convolution layers of different kernel sizes arranged in parallel, and an output convolution module; the three convolution layers with different kernel sizes are arranged in parallel, and the feature images extracted by the convolution layers are spliced with the feature images output by the input convolution module and then output by the output convolution module.
5. The method for detecting internal defects of a gas-insulated switchgear according to claim 2, wherein the neck network based on a transducer network specifically comprises:
the device comprises a convolution layer, a fourth CSPLAyer module, a first mask convolution module, a second mask convolution module, a fifth CSPLAyer module, a sixth CSPLAyer module, a seventh CSPLAyer module and a transducer module;
inputting a third characteristic diagram into the convolution layer, and obtaining a fourth characteristic diagram after convolution operation;
the fourth feature map is spliced with a second feature map input into a fourth CSPLayer module after up-sampling operation, feature map information is fused, and further features are extracted to obtain a first fused feature map;
inputting a first fusion feature map to the first mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a first mask feature map;
inputting a first feature map to the second mask convolution module, and extracting features of the feature map which is shielded according to a distribution rule to obtain a second mask feature map;
the first mask feature map is spliced with the second mask feature map after up-sampling operation, feature map information is fused, a second fused feature map is further obtained through feature extraction of a fifth CSPLayer module, and the second fused feature map is output to a prediction network;
the second fusion feature map is spliced with the first mask feature map after downsampling operation, feature map information is fused, a third fusion feature map is further obtained through feature extraction of a sixth CSPLlayer module, and the third fusion feature map is output to a prediction network;
the third fusion feature map is spliced with the fourth feature map after downsampling operation, feature map information is fused, the fourth fusion feature map is further obtained through feature extraction of a seventh CSPLayer module, and the fourth fusion feature map is output to a transducer module;
and the converter module performs global feature extraction on the fourth fusion feature map to obtain a fifth fusion feature map, and outputs the fifth fusion feature map to the prediction network.
6. The method for detecting internal defects of a gas-insulated switchgear according to claim 5, wherein the first mask convolution module and the second mask convolution module have the same structure, and specifically comprise:
a first mask convolution layer, a second mask convolution layer, a third mask convolution layer, a fourth mask convolution layer, a fifth mask convolution layer, and a sixth mask convolution layer in parallel;
the first mask convolution layer processes the input first fusion feature map by adopting a blank first mask, and then performs feature extraction through the convolution layer and the normalization layer to obtain a first mask convolution feature map;
and the second to sixth mask convolution layers process the input first fusion feature images by adopting the second to sixth masks which are shielded according to a distribution rule, and then perform feature extraction through the convolution layers and the normalization layers to obtain second to sixth mask convolution feature images.
7. The method for detecting internal defects of a gas-insulated switchgear according to claim 5, wherein the transducer module specifically comprises:
a convolutional layer a, transformer network, convolutional layer b, and convolutional layer c;
the convolution layer a performs position coding on the input fourth fusion feature map, performs global feature extraction through a Transformer network, and performs convolution operation through the convolution layer b to obtain a global feature map;
and splicing the global feature map with the fourth fusion feature map through jump connection, and then performing convolution operation through a convolution layer c to obtain a fifth fusion feature map.
8. The method for detecting internal defects of a gas-insulated switchgear according to claim 5 or 6, wherein in the trained REP-YOLOX target detection model based on a convolutional neural network and a transform network, a first mask convolution module is subjected to model weight parameters, an input first fusion feature map is superimposed according to a channel dimension, and six convolutions of the first mask convolution module are superimposed according to the channel dimension.
9. The method for detecting internal defects of a gas-insulated switchgear according to claim 5 or 6, wherein in the trained REP-YOLOX target detection model based on a convolutional neural network and a transform network, a second mask convolution module is subjected to model weight parameters, an input first feature map is superimposed according to a channel dimension, and six convolutions of the second mask convolution module are superimposed according to the channel dimension.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311128356.9A CN117237286B (en) | 2023-09-02 | Method for detecting internal defects of gas-insulated switchgear |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202311128356.9A CN117237286B (en) | 2023-09-02 | Method for detecting internal defects of gas-insulated switchgear |
Publications (2)
Publication Number | Publication Date |
---|---|
CN117237286A true CN117237286A (en) | 2023-12-15 |
CN117237286B CN117237286B (en) | 2024-05-17 |
Family
ID=
Citations (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110827278A (en) * | 2019-11-27 | 2020-02-21 | 云南电网有限责任公司电力科学研究院 | Transformer substation insulator defect online detection method and device |
CN111598843A (en) * | 2020-04-24 | 2020-08-28 | 国电南瑞科技股份有限公司 | Power transformer respirator target defect detection method based on deep learning |
CN112184746A (en) * | 2020-08-27 | 2021-01-05 | 西北工业大学 | Transmission line insulator defect analysis method |
CN114240878A (en) * | 2021-12-16 | 2022-03-25 | 国网河南省电力公司电力科学研究院 | Routing inspection scene-oriented insulator defect detection neural network construction and optimization method |
CN114283117A (en) * | 2021-11-24 | 2022-04-05 | 广西大学 | Insulator defect detection method based on improved YOLOv3 convolutional neural network |
CN114399628A (en) * | 2021-12-21 | 2022-04-26 | 四川大学 | Insulator high-efficiency detection system under complex space environment |
CN114419413A (en) * | 2022-01-12 | 2022-04-29 | 国网河南省电力公司电力科学研究院 | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network |
WO2022126377A1 (en) * | 2020-12-15 | 2022-06-23 | 中国科学院深圳先进技术研究院 | Traffic lane line detection method and apparatus, and terminal device and readable storage medium |
CN115829999A (en) * | 2022-12-22 | 2023-03-21 | 国网新疆电力有限公司信息通信公司 | Insulator defect detection model generation method, device, equipment and storage medium |
CN115953408A (en) * | 2023-03-15 | 2023-04-11 | 国网江西省电力有限公司电力科学研究院 | YOLOv 7-based lightning arrester surface defect detection method |
CN116009581A (en) * | 2022-12-20 | 2023-04-25 | 国网河北省电力有限公司电力科学研究院 | Unmanned aerial vehicle inspection method for power transmission line, unmanned aerial vehicle control terminal and storage medium |
CN116188433A (en) * | 2023-02-28 | 2023-05-30 | 广东电网有限责任公司 | Insulator defect detection method, device and equipment |
US20230184927A1 (en) * | 2021-12-15 | 2023-06-15 | Anhui University | Contextual visual-based sar target detection method and apparatus, and storage medium |
CN116310718A (en) * | 2023-02-03 | 2023-06-23 | 广东工业大学 | Method, system and equipment for detecting pest target based on YOLOv5 model |
CN116416237A (en) * | 2023-04-13 | 2023-07-11 | 广东电网有限责任公司 | Power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement |
Patent Citations (15)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110827278A (en) * | 2019-11-27 | 2020-02-21 | 云南电网有限责任公司电力科学研究院 | Transformer substation insulator defect online detection method and device |
CN111598843A (en) * | 2020-04-24 | 2020-08-28 | 国电南瑞科技股份有限公司 | Power transformer respirator target defect detection method based on deep learning |
CN112184746A (en) * | 2020-08-27 | 2021-01-05 | 西北工业大学 | Transmission line insulator defect analysis method |
WO2022126377A1 (en) * | 2020-12-15 | 2022-06-23 | 中国科学院深圳先进技术研究院 | Traffic lane line detection method and apparatus, and terminal device and readable storage medium |
CN114283117A (en) * | 2021-11-24 | 2022-04-05 | 广西大学 | Insulator defect detection method based on improved YOLOv3 convolutional neural network |
US20230184927A1 (en) * | 2021-12-15 | 2023-06-15 | Anhui University | Contextual visual-based sar target detection method and apparatus, and storage medium |
CN114240878A (en) * | 2021-12-16 | 2022-03-25 | 国网河南省电力公司电力科学研究院 | Routing inspection scene-oriented insulator defect detection neural network construction and optimization method |
CN114399628A (en) * | 2021-12-21 | 2022-04-26 | 四川大学 | Insulator high-efficiency detection system under complex space environment |
CN114419413A (en) * | 2022-01-12 | 2022-04-29 | 国网河南省电力公司电力科学研究院 | Method for constructing sensing field self-adaptive transformer substation insulator defect detection neural network |
CN116009581A (en) * | 2022-12-20 | 2023-04-25 | 国网河北省电力有限公司电力科学研究院 | Unmanned aerial vehicle inspection method for power transmission line, unmanned aerial vehicle control terminal and storage medium |
CN115829999A (en) * | 2022-12-22 | 2023-03-21 | 国网新疆电力有限公司信息通信公司 | Insulator defect detection model generation method, device, equipment and storage medium |
CN116310718A (en) * | 2023-02-03 | 2023-06-23 | 广东工业大学 | Method, system and equipment for detecting pest target based on YOLOv5 model |
CN116188433A (en) * | 2023-02-28 | 2023-05-30 | 广东电网有限责任公司 | Insulator defect detection method, device and equipment |
CN115953408A (en) * | 2023-03-15 | 2023-04-11 | 国网江西省电力有限公司电力科学研究院 | YOLOv 7-based lightning arrester surface defect detection method |
CN116416237A (en) * | 2023-04-13 | 2023-07-11 | 广东电网有限责任公司 | Power transmission line defect detection method based on improved YOLOv5 and fuzzy image enhancement |
Non-Patent Citations (4)
Title |
---|
XUELI XIE 等: "STFTrack: Spatio-Temporal-Focused Siamese Network for Infrared UAV Tracking", 《DRONES》, 28 April 2023 (2023-04-28), pages 1 - 22 * |
刘亚琳: "基于深度学习的车辆目标检测方法研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅱ辑》, no. 11, 15 November 2022 (2022-11-15), pages 034 - 82 * |
张红民 等: "优化YOLO网络的人体异常行为检测方法", 《计算机工程与应用》, no. 7, 30 April 2023 (2023-04-30), pages 242 - 249 * |
王珏: "一种基于深度学习的服装图像实例分割算法研究", 《中国优秀硕士学位论文全文数据库工程科技Ⅰ辑》, no. 01, 15 January 2023 (2023-01-15), pages 024 - 1806 * |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Chen et al. | Fault detection of insulators using second-order fully convolutional network model | |
CN110335260B (en) | Power cable damage detection method based on light convolutional neural network | |
Gao et al. | Efficient parallel branch network with multi-scale feature fusion for real-time overhead power line segmentation | |
CN114782734A (en) | Visual detection method for pipeline leakage of valve cooling system, computer and storage medium | |
CN112750125B (en) | Glass insulator piece positioning method based on end-to-end key point detection | |
CN111079645A (en) | Insulator self-explosion identification method based on AlexNet network | |
Pei et al. | UV‐flashover evaluation of porcelain insulators based on deep learning | |
US11948078B2 (en) | Joint representation learning from images and text | |
Yi et al. | Insulator and defect detection model based on improved YOLO-S | |
CN117237286B (en) | Method for detecting internal defects of gas-insulated switchgear | |
Wang et al. | Insulator anomaly detection method based on few-shot learning | |
CN108761263B (en) | Fault diagnosis system based on evidence theory | |
Wang et al. | Semantic segmentation and defect detection of aerial insulators of transmission lines | |
Jing et al. | SmokeSeger: a transformer-CNN coupled model for urban scene smoke segmentation | |
CN114332047A (en) | Construction method and application of surface defect detection model | |
CN117237286A (en) | Method for detecting internal defects of gas-insulated switchgear | |
CN117197571A (en) | Photovoltaic module fault detection method and device, electronic equipment and storage medium | |
CN116612481A (en) | Knowledge graph and multi-element image-based power equipment defect identification method and system | |
Huang et al. | Research on intelligent detection of sulfur hexafluoride gas leakage in confined spaces | |
CN111144420B (en) | Electric tower defect monitoring method and system based on small samples | |
CN112529881B (en) | Power control cabinet cable anomaly identification method and device | |
CN114862803A (en) | Industrial image anomaly detection method based on fast Fourier convolution | |
CN112966788A (en) | Power transmission line spacer fault detection method based on deep learning | |
CN117113066B (en) | Transmission line insulator defect detection method based on computer vision | |
CN111797925B (en) | Visual image classification method and device for power system |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant |